Lines Matching refs:__m128i
19 __m128i *const d) {
27 __m128i *const d) {
33 __m128i *const d) {
34 d[0] = _mm_loadl_epi64((const __m128i *)(s + 0 * stride));
35 d[1] = _mm_loadl_epi64((const __m128i *)(s + 1 * stride));
36 d[2] = _mm_loadl_epi64((const __m128i *)(s + 2 * stride));
37 d[3] = _mm_loadl_epi64((const __m128i *)(s + 3 * stride));
41 __m128i *const d) {
47 const ptrdiff_t stride, __m128i *const d) {
48 d[0] = _mm_load_si128((const __m128i *)(s + 0 * stride));
49 d[1] = _mm_load_si128((const __m128i *)(s + 1 * stride));
50 d[2] = _mm_load_si128((const __m128i *)(s + 2 * stride));
51 d[3] = _mm_load_si128((const __m128i *)(s + 3 * stride));
52 d[4] = _mm_load_si128((const __m128i *)(s + 4 * stride));
53 d[5] = _mm_load_si128((const __m128i *)(s + 5 * stride));
54 d[6] = _mm_load_si128((const __m128i *)(s + 6 * stride));
55 d[7] = _mm_load_si128((const __m128i *)(s + 7 * stride));
59 const ptrdiff_t stride, __m128i *const d) {
60 d[0] = _mm_loadu_si128((const __m128i *)(s + 0 * stride));
61 d[1] = _mm_loadu_si128((const __m128i *)(s + 1 * stride));
62 d[2] = _mm_loadu_si128((const __m128i *)(s + 2 * stride));
63 d[3] = _mm_loadu_si128((const __m128i *)(s + 3 * stride));
67 const ptrdiff_t stride, __m128i *const d) {
72 static INLINE void _mm_storeh_epi64(__m128i *const d, const __m128i s) {
76 static INLINE void store_8bit_4x4(const __m128i *const s, uint8_t *const d,
84 static INLINE void store_8bit_4x4_sse2(const __m128i s, uint8_t *const d,
86 __m128i ss[4];
95 static INLINE void store_8bit_8x4_from_16x2(const __m128i *const s,
98 _mm_storel_epi64((__m128i *)(d + 0 * stride), s[0]);
99 _mm_storeh_epi64((__m128i *)(d + 1 * stride), s[0]);
100 _mm_storel_epi64((__m128i *)(d + 2 * stride), s[1]);
101 _mm_storeh_epi64((__m128i *)(d + 3 * stride), s[1]);
104 static INLINE void store_8bit_8x8(const __m128i *const s, uint8_t *const d,
106 _mm_storel_epi64((__m128i *)(d + 0 * stride), s[0]);
107 _mm_storel_epi64((__m128i *)(d + 1 * stride), s[1]);
108 _mm_storel_epi64((__m128i *)(d + 2 * stride), s[2]);
109 _mm_storel_epi64((__m128i *)(d + 3 * stride), s[3]);
110 _mm_storel_epi64((__m128i *)(d + 4 * stride), s[4]);
111 _mm_storel_epi64((__m128i *)(d + 5 * stride), s[5]);
112 _mm_storel_epi64((__m128i *)(d + 6 * stride), s[6]);
113 _mm_storel_epi64((__m128i *)(d + 7 * stride), s[7]);
116 static INLINE void storeu_8bit_16x4(const __m128i *const s, uint8_t *const d,
118 _mm_storeu_si128((__m128i *)(d + 0 * stride), s[0]);
119 _mm_storeu_si128((__m128i *)(d + 1 * stride), s[1]);
120 _mm_storeu_si128((__m128i *)(d + 2 * stride), s[2]);
121 _mm_storeu_si128((__m128i *)(d + 3 * stride), s[3]);
